def __init__(self, ntoken, ninp, nhid, nhidlast,
dropout=0.5, dropouth=0.5, dropoutx=0.5, dropouti=0.5, dropoute=0.1,
cell_cls=DARTSCell, genotype=None):
super(RNNModel, self).__init__()
self.lockdrop = LockedDropout()
self.encoder = nn.Embedding(ntoken, ninp)
assert ninp == nhid == nhidlast
if cell_cls == DARTSCell:
assert genotype is not None
self.rnns = [cell_cls(ninp, nhid, dropouth, dropoutx, genotype)]
else:
assert genotype is None
self.rnns = [cell_cls(ninp, nhid, dropouth, dropoutx)]
self.rnns = torch.nn.ModuleList(self.rnns)
self.decoder = nn.Linear(ninp, ntoken)
self.decoder.weight = self.encoder.weight # weight tying
self.init_weights()
self.ninp = ninp
self.nhid = nhid
self.nhidlast = nhidlast
self.dropout = dropout
self.dropouti = dropouti
self.dropoute = dropoute
self.ntoken = ntoken
self.cell_cls = cell_cls
def __init__(self,
n_token,
n_inp,
n_hid,
n_hid_last,
dropout=0.5,
dropout_h=0.5,
dropout_x=0.5,
dropout_i=0.5,
dropout_e=0.1,
cell_cls=DARTSCell):
super().__init__()
self.lockdrop = LockedDropout()
self.encoder = nn.Embedding(
n_token, n_inp) # n_inp 词向量长度 embedding的作用是把输入的index转化为向量
self.rnn = cell_cls(n_inp, n_hid, dropout_h, dropout_x)
self.decoder = nn.Linear(n_inp, n_token)
self.decoder.weight = self.encoder.weight
self.init_weights()
self.n_inp = n_inp
self.n_hid = n_hid
self.n_hid_last = n_hid_last
self.dropout = dropout
self.dropout_i = dropout_i
self.dropout_e = dropout_e
self.n_token = n_token
self.cell_cls = cell_cls
def __init__(self, ntoken, args, cell_cls=DARTSCell, genotype=None):
super(RNNModel, self).__init__()
self.lockdrop = LockedDropout()
self.encoder = nn.Embedding(ntoken, args.emsize)
self.nlayers = args.nlayers
self.rnns = []
assert args.emsize == args.nhid == args.nhidlast
for layers in range(args.nlayers):
if cell_cls == DARTSCell:
assert genotype is not None
self.rnns.append(
cell_cls(args.emsize, args.nhid, args.dropouth,
args.dropoutx, genotype,
args.use_matrices_on_edge, args.use_glorot,
args.num_intermediate_nodes,
args.handle_hidden_mode))
else:
assert genotype is None
# TODO: for the init of DartsCell there is no default value for genotype,
# next line will raise error in theory
self.rnns = [
cell_cls(args.emsize, args.nhid, args.dropouth,
args.dropoutx)
]
self.rnns = torch.nn.ModuleList(self.rnns)
self.decoder = nn.Linear(args.emsize, ntoken)
self.decoder.weight = self.encoder.weight
self.init_weights()
self.ninp = args.emsize
self.nhid = args.nhid
self.nhidlast = args.nhidlast
self.dropout = args.dropout
self.dropouti = args.dropouti
self.dropoute = args.dropoute
self.ntoken = ntoken
self.cell_cls = cell_cls
model_logger.info("MODEL INITIALIZED ")
def __init__(self,
ntoken,
ninp,
nhid,
nhidlast,
dropout=0.5,
dropouth=0.5,
dropoutx=0.5,
dropouti=0.5,
dropoute=0.1,
nner=None,
npos=None,
token_emb_path=None,
nclasses=None,
nloc=None,
cell_cls=DARTSCell,
genotype=None):
super(RNNModel, self).__init__()
# TACRED attributes
self.nner = nner
self.npos = npos
self.nhid = nhid
self.nloc = nloc
self.token_emb_path = token_emb_path
self.nclasses = nclasses
# Original attributes
self.ninp = ninp
self.nhidlast = nhidlast
self.dropout = dropout
self.dropouti = dropouti
self.dropoute = dropoute
self.ntoken = ntoken
self.cell_cls = cell_cls
self.lockdrop = LockedDropout()
self.encoder = nn.Embedding(ntoken, ninp, padding_idx=constant.PAD_ID)
self.peripheral_emb_dim = 0
if self.nner is not None:
self.ner_encoder = nn.Embedding(len(constant.NER_TO_ID),
self.nner,
padding_idx=constant.PAD_ID)
self.peripheral_emb_dim += self.nner
if self.npos is not None:
self.npos_encoder = nn.Embedding(len(constant.POS_TO_ID),
self.npos,
padding_idx=constant.PAD_ID)
self.peripheral_emb_dim += self.npos
if self.nloc is not None:
self.nloc_encoder = nn.Embedding(constant.MAX_LEN * 2 + 1,
self.nloc)
self.peripheral_emb_dim += self.nloc + self.nloc
# If using additional token attributes, need to encode them using smaller size
if self.peripheral_emb_dim > 0:
input_dim = self.ninp + self.peripheral_emb_dim
self.input_aggregator = nn.Linear(in_features=input_dim,
out_features=self.ninp)
assert ninp == nhid == nhidlast
if cell_cls == DARTSCell:
assert genotype is not None
self.rnns = [cell_cls(ninp, nhid, dropouth, dropoutx, genotype)]
else:
assert genotype is None
self.rnns = [cell_cls(ninp, nhid, dropouth, dropoutx)]
self.rnns = torch.nn.ModuleList(self.rnns)
self.decoder = nn.Linear(ninp, nclasses)
# self.decoder.weight = self.encoder.weight
self.init_weights()